Cython 具有不同类型的相同类

Question

I have two cdef classes B and C whose methods are exactly the same. The only difference between them is the type of their attributes: one has mpz attributes, the other one int attributes.

My first guess was to use an abstract class A which would be overriden by B and C . The problem is that Cython apparently doesn't want me to override attributes (plus, which type should I give to the attribute of the abstract class?). The error I got by doing this was:

------------------------------------------------------------
...

cdef class TestModularNumber:
    cdef readonly mpz value, modulo

cdef class TestInheritance(TestModularNumber):
    cdef readonly int value, modulo                     ^
------------------------------------------------------------

finite_field/testmodular.pxd:10:22: 'value' redeclared

My second guess was to use a single class with fused type like this:

ctypedef fused mpz_or_int:
    int
    mpz

But Cython complains about the fact that I do operations with this type (like % , even though it is defined for both of the types). The errors I got are:

------------------------------------------------------------
...
ctypedef fused mpz_or_int:
    int
    mpz

cdef class TestModularNumber:
    cdef readonly mpz_or_int value, modulo                                   ^
------------------------------------------------------------

finite_field/testmodular.pxd:12:36: Type is not specialized

------------------------------------------------------------
...
from gmpy2 import invert, powmod


cdef class TestModularNumber:
    def __cinit__(self, mpz_or_int value, mpz_or_int modulo):
        self.value = value
           ^
------------------------------------------------------------

finite_field/testmodular.pyx:7:12: Invalid use of fused types, type cannot be specialized

------------------------------------------------------------
...
        if not self.has_modular_square_root():
            raise ValueError(f"{self.value} is a non-residue modulo {self.modulo}.")
        if self.value == 0 or self.value == 1:
            return TestModularNumber(self.value, self.modulo)

        if self.modulo % 4 == 3:
                      ^
------------------------------------------------------------

finite_field/testmodular.pyx:145:23: Compiler crash in AnalyseExpressionsTransform

For now, I copy/pasted the two classes, but that's a dirty hack, and obviously, every time that I have to modify a method of these classes, I cry:)

I think the way to go is using fused types, but how can I find a workaround to this problem?

Answer 1

For reasons explained in the Cython: templates in python class wrappers this generally isn't available in Cython. cdef classes can't have fused members. In the answer to that question I proposed using the "copy/paste" method, but maybe trying to automate it.

For this specific case I wonder if you could do a bit better by taking the implementation out of the class into separate cdef fused functions. First define a fused class for your cdef classes

ctypedef fused cdef_pz_or_int:
    # cdef classes _can_ be part of fused types
    TestModularNumber
    TestInheritance

I'd recommend not having any inheritance relationship between the classes - just have them implement a (semi-)common interface.

The actual implementation is then in non-member fused functions:

cdef has_modular_square_root(cdef_pz_or_int self):
    value = self.value  # cython should be able to infer this type
    # more logic goes here....

You could obviously mix and match cdef_pz_or_int and pz_or_int as arguments (if you need to do so) however it'd probably need to generate valid code for all combinations.

---

I'm not 100% confident this will work for you - if not I'm happy to delete it.